Original title: KRAFTON/Raon-1-nano
A dense small model from KRAFTON AI, a subsidiary of KRAFTON, the developer of PUBG: effective parameters 1.5B, total parameters 3.1B (including a 1.5B per-layer embedding table), context 262,144 tokens, thinking mode enabled by default. Out of 24 layers, 18 layers use linear attention (Raon Delta Attention), and only 6 layers of MLA retain KV cache, approximately 7 KiB per token. It was derived from Raon-1 mini preview (30B-A4B) via MoE-to-dense pruning and distillation, followed by continued pre-training on 3T tokens, then SFT, online distillation, and reinforcement learning. Official comparisons with peer models such as Qwen3.5 2B and Gemma 4 E2B show it achieving the highest scores in the table for IFBench 75.9, FACTS Grounding 44.4, and KMMLU-Redux 50.9, while BFCL v4 scores 56.3, lower than LFM3 1.7B's 67.2.
Why it matters Another example of a non-English localized small model; official GGUF, MLX, and LiteRT-LM versions are provided, allowing local execution on laptops and Apple silicon devices
Upvotes 48 · Likes 48 · Downloads 58
Reasons
- Trained bilingually in Korean and English, capable of calling tools during conversations; official tests cover office tasks, mobile device control, terminal commands, and search
- Comprehensive format support: official versions available for GGUF, MLX (4-bit), and LiteRT-LM, enabling local execution on Apple silicon
- Weights licensed under Apache-2.0, allowing commercial use
Usage
- llama.cpp: Compile KRAFTON's raon-1-nano-preview branch, then run llama-cli -hf KRAFTON/Raon-1-nano-gguf:Q4_K_M --reasoning on
- Apple silicon: Install mlx==0.32.4 and mlx-lm via pip, then load KRAFTON/Raon-1-nano-mlx (trust_remote_code=True)
- Official recommendations: For daily conversation, use temperature 0.1 and presence_penalty 1.5; for agents and code, use temperature 1.0
Risks
- vLLM and llama.cpp currently require KRAFTON's own branches; mainline support has not yet been merged
- Office Work Bench and Chatbot Korean are unpublished internal tests; the model card does not specify Chinese language capabilities
- The transformers version processes prompts token-by-token, resulting in slow prefilling for long prompts; the LiteRT-LM version has only a 4,096 token context and does not support thinking mode
- This entry has not been tested; refer to the model card for performance details